Knowledge Enhanced Sentiment Analysis

As a crucial subtask in natural language processing,sentiment analysis has received great attention from academia and industry.This task aims to predict users' emotions,sentiments,opinions,and attitudes in a given text.With the development of Internet,many forums(e.g.,blogs,social networks,ecommerce sites,news reports,and other online resources)can be used to express opinions.Sentiment analysis on these texts can help understand the public and consumers' opinions on social events,political movements,corporate strategies,marketing activities,and product preferences to monitor public opinion and improve company products and services.Previous researches on sentiment analysis are mainly rulebased,lexiconbased,and machine learningbased methods.Recently,with the success of neural networks in various natural language processing tasks,deep learning has also been widely used in sentiment analysis tasks and achieved excellent performance.The current sentiment analysis system is datadriven,which trains neural networks based on labeled data to capture the explicit sentiment by text matching.Most of the existing studies treat the sentiment analysis task as a special text classification task and ignore its characteristics,including ambiguity and implicitness.There are still some unsolved problems: 1)the sentiment expressions in different domains are also inconsistent,which leads to poor performance of transfer learning? 2)the current sentiment analysis data sets are too small to learn domain sentiment knowledge,which greatly limits the deep learning model's effectiveness? 3)some neutral words(sentiment commonsense knowledge)that express different sentiments in different contexts(e.g.,“fast”and “hot”)are ignored by the current work? 4)most of the current studies focus on shallow sentiment recognition,and deep sentiment cannot be mined by reasoning based on sentiment knowledge.The main reason for these problems is that the existing models lack domain and commonsense sentiment knowledge.Therefore,in view of the above shortcomings in sentiment analysis tasks,this article conducts indepth explorations from pretraining,transfer learning,unsupervised sentiment commonsense extraction,and sentiment reasoning to integrate the knowledge implicitly and explicitly,and improve the models' accuracy.Specifically,the main contributions of this article are summarized as follows:(1)For the first question above,we propose a pretraining method to model crossdomain sentiment knowledge implicitly.We propose to learn the domaininvariant sentiment knowledge in the pretraining stage for crossdomain sentiment analysis.We propose a sentimentaware masking language model and sentimentaware pretraining goals in word and sentence levels to learn the sentiment knowledge contained in the text.Experiments show that the pretraining model can greatly improve the model's accuracy in crossdomain sentiment analysis tasks.(2)For the second question above,we propose a hierarchical transfer method to model indomain sentiment knowledge implicitly.We propose to transfer resourcerich sentencelevel sentiment knowledge to the aspectlevel sentiment analysis task.We pretrain the model based on the sampled sentencelevel example and finetune the model on the aspectlevel sentiment analysis task data set.We also explore the influnce of each layer.A series of experiments have proved that this transfer method can greatly improve the accuracy of aspectlevel sentiment analysis.(3)For the third question above,we propose an unsuperivised method to extract sentiment extraction explicitly.We design an unsupervised way to extract the sentiment knowledge contained in the text.For example,“fast” expresses a positive sentiment for logistics and expresses a negative sentiment for power consumption.Discrete and continuousbased perturbations are proposed to extract the opinion words corresponding to the aspect.Both manual evaluation and automatic evaluation show that the model in this paper can effectively extract the sentiment knowledge contained in the text.(4)For the fourth question above,we propose a graph reasoning method to model sentiment extraction explicitly.We introduce a sentiment knowledge and dependency based reasoning model for finegrained sentiment analysis.A graph neural network is designed to model the syntactic information and sentiment knowledge independently or jointly for graph reasoning.